Wind power long-term power generation prediction method based on time-varying copula-gbdt

Abstract

The application discloses a wind power long-term power generation prediction method based on a time-varying Copula-GBDT, and relates to the field of wind farm power generation prediction. Wind farm measured data and historical reanalysis meteorological data are collected and data preprocessing is carried out; the wind shear power law formula is taken as a physical constraint and is integrated into a GRU neural network model for correction; a time-varying t-Copula model is constructed, and key meteorological variables are screened; the corrected wind speed is fitted by using a Weibull distribution, the wind speed interval is integrated, and then the improved LSTM power curve is input, and the monthly power generation is accumulated; the GBDT model is constructed by taking the power generation of the same month in previous years and the screened key meteorological variables as input characteristics, a dynamic elastic weight regularization is introduced, and the monthly power generation in a future preset time period, annual total power generation and a prediction trend curve are output. The problems of sample sparsity, low meteorological data correction accuracy, insufficient nonlinear feature mining and weak climate drift resistance are solved.

Description

Technical Field

[0001] This invention relates to the field of wind farm power generation prediction, and more specifically, to a long-term wind power generation prediction method based on time-varying Copula-GBDT. Background Technology

[0002] Long-term wind power generation forecasting is a crucial supporting technology for new energy planning, grid dispatching, and electricity trading. However, current domestic and international long-term wind power generation forecasting technologies still have many shortcomings and are difficult to adapt to complex application scenarios.

[0003] Currently, long-term wind power generation forecasts mostly employ traditional statistical methods such as the persistence method, linear regression, and ARIMA. However, existing forecasts can only extrapolate short-term data and cannot capture long-term climate trends. Some machine learning models suffer from insufficient generalization ability due to the sparseness of monthly samples, resulting in an average prediction accuracy of less than 87.5%. Furthermore, there is a lack of site-level reanalysis meteorological data correction methods; wind speed correction models such as GRU do not incorporate physical mechanisms, leading to poor physical consistency; feature selection relies on linear correlation analysis, which cannot characterize nonlinearity and tail correlation between variables, and the fixed regularization coefficient of traditional EWC makes it difficult to adapt to seasonal and abrupt climate change scenarios.

[0004] In summary, existing technologies suffer from common problems such as sparse samples, low accuracy of meteorological data correction, insufficient nonlinear feature mining, and weak resistance to climate drift, resulting in low prediction accuracy and poor stability. Summary of the Invention

[0005] In view of this, the present invention provides a long-term wind power generation prediction method based on time-varying Copula-GBDT, which solves the problems of sparse samples, low accuracy of meteorological data correction, insufficient nonlinear feature mining, and weak resistance to climate drift in existing long-term wind power generation prediction methods.

[0006] To achieve the above objectives, the following solution is proposed: The long-term wind power generation prediction method based on time-varying Copula-GBDT includes: Collect measured data from wind farms and historical reanalysis meteorological data, and perform anomaly identification, missing data imputation and restoration to obtain multi-source datasets; Using historical measured wind speeds as supervised values, the wind shear power law formula is incorporated into the GRU neural network model as a physical constraint to perform localized correction on the reanalysis wind speed. We constructed a time-varying t-Copula model that distinguishes by month, independently fitted the tail correlation parameters month by month, and combined the rank correlation coefficient to explore the correlation between variables and screen key meteorological variables. The wind speed is fitted with a Weibull distribution and corrected. The wind speed interval is integrated and then input into the improved LSTM power curve to convert it into hourly theoretical power. The monthly power generation is then accumulated. A GBDT model was constructed using the power generation of the same month in previous years and the selected key meteorological variables as input features, and dynamic elastic weight regularization was introduced. The GBDT model outputs monthly power generation, annual total power generation, and predicted trend curves for a future preset time period.

[0007] Preferably, it includes: The three-sigma principle, bidirectional quartile method and DBSCAN clustering are used to jointly identify abnormal data; By combining KNN with LSTM for temporal extrapolation, missing data can be imputed and restored to obtain a multi-source dataset.

[0008] Preferably, the total loss function of the GRU neural network model is: ; in, This represents the mean square error of wind speed prediction. For physical prior weights, Output wind speed for the GRU neural network model. This is for calculating wind speed based on the power-law formula for wind shear.

[0009] Preferably, the time-varying t-Copula model, which independently fits the tail correlation parameters month by month and combines the rank correlation coefficient to mine the association between variables, includes: Based on multi-meteorological variable data from a multi-source dataset, a time-varying t-Copula model is constructed by month, and the correlation matrix and degree of freedom parameters are fitted month by month. Traverse all meteorological variables and extract the binary correlation coefficients between each set of meteorological variables and power generation from the correlation matrix; The Kendall rank correlation coefficient and Spearman rank correlation coefficient between each meteorological variable and power generation were calculated using the t-Copula function based on each binary correlation coefficient. Determine the association between variables based on the correlation coefficient threshold; Key meteorological variables were selected based on correlations.

[0010] Preferably, the process of screening key meteorological variables includes: If the absolute value of the Kendall rank correlation coefficient of the current meteorological variable is greater than the Kendall coefficient threshold, and the absolute value of the Spearman rank correlation coefficient is greater than the Spearman coefficient threshold, then the current meteorological variable is a key meteorological variable. If the absolute value of the rank correlation coefficient of the current meteorological variable only meets the Kendall coefficient threshold or the Spearman coefficient threshold, then the meteorological variable with direct physical effect is retained as the key meteorological variable based on the physical mechanism. If the absolute value of the rank correlation coefficient of the current meteorological variable does not meet the Kendall coefficient threshold and the Spearman coefficient threshold, then the current meteorological variable will be removed.

[0011] Preferably, the probability density formula for the t-Copula function is: ; in, As a meteorological variable dimension, For the first m The correlation matrix obtained by fitting over a month For the first m The degrees of freedom obtained by fitting over a month For the first i The empirical distribution function values ​​of the meteorological variables It follows a univariate t-distribution. Single variable t The inverse function of the distribution; The formula for Kendall's rank correlation coefficient is: ; in, This represents the binary correlation coefficient between meteorological variables and power generation within the monthly correlation matrix. The Kendall rank correlation coefficient between the two. The formula for the Spearman rank correlation coefficient is: .

[0012] Preferably, wind speed is converted into hourly theoretical power based on an improved LSTM power curve: ; in, For LSTM model parameters, This refers to the rated power of the fan. This is the corrected wind speed; when the wind speed is lower than the cut-in wind speed or higher than the cut-out wind speed. ; The monthly power generation is obtained by summing them up: ; Where N is the number of hours in the current month. for t Wind speed after constant correction For time intervals.

[0013] Preferably, the prediction formula of the GBDT model is: ; in, For the first m The predicted values ​​of each decision tree. For learning rate, For the first m The output of each decision tree; The calculation process for the dynamic elastic weight regularization coefficient is as follows: ; in, As the benchmark coefficient, For adaptive coefficients, For the rate of change of prediction error; The dynamic elastic weight regularization loss function is: ; in, For cross-entropy loss, For dynamic regularization coefficients, For the Fisher information matrix i One element, These are the current model parameters. These are the historically optimal model parameters.

[0014] According to specific embodiments provided by the present invention, the present invention discloses the following technical effects: The wind power long-term generation prediction method based on time-varying Copula-GBDT provided in this invention collects measured wind farm data and historical reanalysis meteorological data, and performs anomaly identification, missing data imputation and restoration. Using historical measured wind speed as a supervised value, the wind shear power-law formula is incorporated as a physical constraint into the GRU neural network model to perform localized correction on the reanalysis wind speed. This invention embeds the physical mechanism of atmospheric motion into the GRU neural network, overcoming the limitations of purely data-driven approaches, and the correction results are more consistent with meteorological and physical laws.

[0015] This invention constructs a time-varying t-Copula model differentiated by month, independently fitting tail correlation parameters month by month, and combining rank correlation coefficients to mine variable associations and screen key meteorological variables. This invention overcomes the limitation of static Copula models in adapting to seasonal variations by independently fitting tail correlation parameters month by month, dynamically capturing the nonlinear correlation changes between different climatic period characteristics and power generation, adapting to different climatic conditions. After feature screening, the model's computational load is significantly reduced, and prediction accuracy is further improved.

[0016] This invention uses a Weibull distribution to fit and correct the wind speed, integrates the wind speed interval, and inputs it into an improved LSTM power curve to convert it into hourly theoretical power, which is then accumulated to obtain the monthly power generation. This invention achieves complete sample expansion based on GRU correction of reanalysis data, Weibull fitting, and improved power curve conversion. Combining Weibull distribution fitting and monthly power generation accumulation, it can expand the training sample by 30-50 times, significantly improving the model's generalization ability and avoiding underfitting problems.

[0017] This invention constructs a GBDT model using historical monthly power generation and selected key meteorological variables as input features, and introduces dynamic elastic weight regularization. The GBDT model of this invention incorporates dynamic elastic weight regularization, with the regularization coefficient adjusted in real time according to the prediction error. It strengthens the protection of historical knowledge during abrupt climate changes and allows the model to adapt to new distributions during stable periods, thus better adapting to seasonal and climate abrupt changes, improving the model's resistance to climate drift, and significantly enhancing model stability. Attached Figure Description

[0018] To more clearly illustrate the technical solutions in the embodiments of the present invention or the prior art, the drawings used in the description of the embodiments or the prior art will be briefly introduced below. Obviously, the drawings described below are only embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on the provided drawings without creative effort.

[0019] Figure 1 A flowchart of a long-term wind power generation prediction method based on time-varying Copula-GBDT provided in an embodiment of the present invention; Figure 2 The monthly power generation prediction comparison curves for typical wind farms provided in the embodiments of the present invention are shown. Detailed Implementation

[0020] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0021] First, combined Figure 1 This invention introduces a long-term wind power generation prediction method based on time-varying Copula-GBDT, as provided in the embodiments of the present invention. Figure 1 As shown, the method includes: Step S1: Collect measured data and historical reanalysis meteorological data of the wind farm, and perform abnormal data identification, missing data imputation and restoration.

[0022] Specifically, the system collects measured power, wind speed, temperature, air pressure, and humidity data from the wind farm. The sampling frequency can be 15 minutes, and the data availability rate is ≥90%. It can also acquire historical reanalysis meteorological data for a period of Q years, where Q can be set to 50. The reanalysis meteorological data includes core meteorological variables such as wind speed, temperature, air pressure, and humidity.

[0023] The three-sigma principle, bidirectional quartile method and DBSCAN clustering are used to jointly identify marginal, clustered and scattered abnormal data, with an anomaly identification accuracy of ≥98.2%.

[0024] (1) The 3σ principle anomaly identification formula is: ; in, The mean of the data. Let be the standard deviation of the data. Data that satisfies this formula are considered outliers.

[0025] (2) The formula for anomaly identification using the bidirectional quartile method is: ,or, ; in, It is the first quartile. It is the third quartile. It is the interquartile range.

[0026] KNN+LSTM temporal extrapolation is used to impute and restore missing wind speed and power data. The wind speed interpolation error RMSE ≤ 0.5 m / s and the power interpolation error RMSE ≤ 2.1 MW are achieved, resulting in a high-quality multi-source dataset.

[0027] Step S2: Using historical measured wind speeds as supervised values, the wind shear power law formula is incorporated into the GRU neural network model as a physical constraint to perform localized correction on the reanalysis wind speed.

[0028] Specifically, using historical measured wind speed as the supervised value, the wind shear power law formula is introduced as a physical prior, and auxiliary inputs or loss regularization terms are incorporated into the GRU neural network model to construct a GRU neural network model that integrates physical mechanisms and data-driven approaches.

[0029] The GRU neural network model consists of an input layer, hidden layers, an output layer, and a supervision signal. The input layer takes into account the Q(50)-year reanalysis wind speed time-series features (including time, wind speed, etc.). The hidden layer contains 32 GRU neurons with the tanh activation function and physical constraint regularization terms. The output layer outputs the wind speed after site-level correction, and the supervision signal is historical measured wind speed data. The GRU neural network model achieves localized correction of the reanalysis wind speed data, eliminating systematic bias and grid representativeness errors.

[0030] (1) The power law formula for wind shear is: ; in, Let z be the wind speed at height z. The wind speed is at a height of 10m, and α is the wind shear index (dynamically adjusted according to topography, season, and atmospheric stability).

[0031] (2) Formula for the structure of the GRU neural network model: Reset Door: ; Update Gate: ; Candidate hidden state: ; Hidden state: ; Where σ is the sigmoid activation function, and ⊙ is the element-wise product. for t Continuously analyze the input characteristics of wind speed. for t The state is hidden at time -1. This is the weight matrix. This is a bias term.

[0032] (3) The total loss function of the GRU neural network model is: ; in, This represents the mean square error of wind speed prediction. For physical prior weights, Output wind speed for the GRU neural network model. This is for calculating wind speed based on the power-law formula for wind shear.

[0033] Step S3: Construct a time-varying t-Copula model differentiated by month, independently fit the tail correlation parameters month by month, combine the rank correlation coefficient to explore the correlation between variables, and screen key meteorological variables.

[0034] Specifically, based on multi-meteorological variable data from a multi-source dataset, a time-varying t-Copula model is constructed, differentiated by month, and the tail correlation matrix is ​​independently fitted month by month. With degrees of freedom parameters Then, iterate through all meteorological variables from the correlation matrix. Extract the binary correlation coefficients between each set of meteorological variables and power generation. ρ Based on the t-Copula function and each binary correlation coefficient... ρKendall's rank correlation coefficients between various meteorological variables and power generation under different seasons and climatic conditions were calculated. τ Spearman rank correlation coefficient ρ s This information is used for subsequent feature selection. Based on correlation coefficient thresholds, the nonlinear, tail-related, and extreme weather-related relationships between meteorological variables such as wind speed, temperature, and air pressure and power generation under different seasons and climatic conditions are dynamically mined. Based on these relationships, key meteorological variables (such as wind speed, temperature, and air pressure) are selected, redundant features are eliminated, and the computational load of the model is reduced.

[0035] (1) The probability density formula for the t-Copula function is: ; in, For feature dimension, For the first m Monthly correlation matrix For the first m One month of freedom For the first i The empirical distribution function values ​​of each feature It follows a univariate t-distribution. Single variable t The inverse function of the distribution.

[0036] (2) The Kendall rank correlation coefficient formula is: ; in, Monthly correlation matrix The binary correlation coefficient between internal meteorological variables and power generation. Let be the Kendall rank correlation coefficient between the two.

[0037] The Kendall rank correlation coefficient formula above is the binary correlation coefficient of the t-Copula model. ρ The theoretical mapping and transformation formula with the Kendall rank correlation coefficient. Furthermore, the Kendall rank correlation coefficient can be calculated using Kendall sample statistics formulas, without relying on the binary correlation parameters calculated based on the t-Copula model, but based on raw sample data such as measured meteorological data and power generation. ; in, For symbolic functions, , For historical meteorological variable samples, , As a sample of historical power generation, n This represents the number of samples.

[0038] (3) The formula for Spearman's rank correlation coefficient is: ; in, The binary correlation coefficient between internal meteorological variables and power generation. Let be the Spearman rank correlation coefficient between the two.

[0039] Similarly, without relying on the binary correlation parameters calculated using the t-Copula model, the Spearman rank correlation coefficient can be calculated based on raw sample data such as measured meteorological data and power generation using the Spearman sample statistics formula. ; in, For the first i The difference between the characteristic rank and the power generation rank of a sample of historical meteorological variables.

[0040] (4) The specific process of determining the association between variables based on the correlation coefficient threshold includes: Based on preset correlation coefficient thresholds, meteorological variables for each month are filtered to obtain a set of exclusive key meteorological variables that fit the seasonal and climatic characteristics of that month. The preset correlation coefficient thresholds include the Kendall coefficient threshold. Spearman coefficient threshold .

[0041] For each meteorological variable in the current month, the Kendall rank correlation coefficient τ Spearman rank correlation coefficient ρ s If the correlation coefficient meets the dual threshold, the meteorological variable is determined to be a strongly correlated variable and retained as a key meteorological variable. If only a single correlation coefficient threshold is met, then physical mechanisms are combined for screening. Meteorological variables that have a direct physical effect on wind power output, such as wind speed, temperature, and air pressure, are retained as key meteorological variables, while secondary meteorological variables without direct physical correlation are eliminated. If any correlation coefficient threshold is not met, the meteorological variable is determined to be a weakly correlated variable and is directly removed.

[0042] The feature dimension reduction rate is: ; in, For the original meteorological variable characteristic dimension, These are the characteristic dimensions of the filtered meteorological variables.

[0043] Feature filtering based on t-Copula function, Kendall rank correlation coefficient and Spearman rank correlation coefficient reduced feature dimension by 45% and improved prediction accuracy by ≥3.0%.

[0044] Step S4: Fit the corrected wind speed using a Weibull distribution, integrate the wind speed interval, input the result into the improved LSTM power curve to convert it into hourly theoretical power, and sum them up to obtain the monthly power generation.

[0045] Specifically, the corrected wind speed is fitted with a Weibull distribution. The wind speed meteorological data is integrated over an interval and then input into the improved LSTM power curve (interval mapping rather than single-point mapping) to convert it into hourly theoretical power. The monthly power generation is accumulated to realize the expansion of the reanalysis data into training samples.

[0046] (1) The probability density formula of the Weibull distribution is: ; in, This is the corrected wind speed. k is the shape parameter, and c is the scale parameter.

[0047] (2) The improved LSTM power curve is as follows: ; in, For LSTM model parameters, This refers to the rated power of the fan. This is the corrected wind speed; when the wind speed is lower than the cut-in wind speed or higher than the cut-out wind speed. .

[0048] (3) The monthly power generation is obtained by summing up: ; Where N is the number of hours in the current month. for t Wind speed after constant correction The time interval is usually 1 hour.

[0049] Weibull distribution goodness of fit R 2 ≥0.96, goodness of fit R 2 The calculation process is as follows: ; in, To measure the wind speed, To fit the wind speed to the Weibull distribution, This represents the average measured wind speed.

[0050] After processing, the sample size is increased by 30 to 50 times compared to the original measured sample. The formula for calculating the sample expansion factor is as follows: ; in, To expand the sample size, This represents the original measured sample size.

[0051] Step S5: Construct a GBDT model using historical power generation data for the same month and selected key meteorological variables as input features, and introduce dynamic elastic weight regularization.

[0052] Specifically, the GBDT model is constructed using the power generation of the same month in previous years and key meteorological variables selected by time-varying t-Copula as input features and the power generation of the current month as output. Dynamic elastic weight (EWC) regularization is introduced, and the regularization coefficient is changed from a fixed value to a dynamic adjustment with the rate of change of prediction error. This enhances the protection of historical knowledge during climate abrupt changes and allows the GBDT model to adapt to new distributions during stable periods, thus suppressing seasonal and interannual concept drift.

[0053] The GBDT model consists of an input layer, hidden layers, a regularization layer, and an output layer. The hidden layer comprises a GBDT decision tree (100 trees, depth 8). The regularization layer includes dynamic EWC elastic weights (adaptively adjusted according to the rate of change of prediction error). The output layer outputs the predicted power generation value for the current month. The GBDT model training convergence speed is ≤150 steps.

[0054] (1) The prediction formula of the GBDT model is: ; in, For the first m The predicted values ​​of each decision tree. For learning rate, For the first m The output of each decision tree.

[0055] (2) The calculation process of the dynamic elastic weight regularization coefficient is as follows: ; : in, As the benchmark coefficient, For adaptive coefficients, To predict the rate of change of error, The prediction error at the current moment, This represents the prediction error from the previous moment.

[0056] (3) The dynamic elastic weight regularized loss function is: ; in, For cross-entropy loss, For dynamic regularization coefficients, For the Fisher information matrix i One element, These are the current model parameters. These are the historically optimal model parameters.

[0057] Step S6: The GBDT model outputs the monthly power generation, annual total power generation, and predicted trend curves for the future preset time period.

[0058] Specifically, it outputs the monthly power generation for the next 12 months, the total annual power generation, and the predicted trend curve, and reports them in e-text format to meet the data reporting requirements of the power grid dispatching agency.

[0059] (1) The formula for calculating the annual forecast power generation is: ; in, Forecast annual power generation, For the first m Projected power generation for the month.

[0060] The formula for calculating prediction accuracy is: ; in, Forecast annual power generation, This represents the annual measured power generation.

[0061] The wind power long-term generation prediction method based on time-varying Copula-GBDT in this invention uses multi-source meteorological data fusion, localized correction of reanalysis data with fused physical priors, time-varying nonlinear correlation mining, and time-series driven machine learning to predict the monthly / annual long-term generation of wind farms. It can adapt to various terrains such as plains, mountains, and hills, as well as wind farms under different climatic conditions. It provides reliable technical support for new energy meteorological support, power grid planning, and wind farm operation decisions. It is applicable to scenarios such as the operational deployment of provincial meteorological service centers, power grid new energy planning, wind farm operation decisions, and power trading.

[0062] Next, the feasibility of the long-term wind power generation prediction method based on time-varying Copula-GBDT is verified through comparative experiments in this embodiment of the invention. The experiment is conducted using data from a typical mountain wind farm in Hebei Province (installed capacity 50MW, measured data from 2024-2025, sampling frequency 15min). The experimental process is as follows: Step S1: Collect measured data and historical reanalysis meteorological data of the wind farm, and perform abnormal data identification, missing data imputation and restoration.

[0063] Data collection: Data collected from this wind farm in 2024 The data includes measured power, wind speed, temperature, air pressure, and humidity data from 2025, as well as 50 years of historical reanalysis meteorological data (including factors such as wind speed, temperature, air pressure, and humidity), ensuring a data availability rate of ≥90%.

[0064] Anomaly data were cleaned using the 3σ principle, bidirectional quartile method, and DBSCAN clustering, and marginal, clustered, and dispersed anomalies were identified, achieving an anomaly identification accuracy of 98.7%. KNN and LSTM were used for time-series extrapolation to impute missing data. Substituting these into the RMSE calculation formula, the wind speed interpolation RMSE was found to be 0.42 m / s, and the power interpolation RMSE was 1.98 MW.

[0065] The combined data preprocessing method of the present invention, which uses the 3σ principle, bidirectional quartiles, DBSCAN, KNN, and LSTM, is compared with a single data preprocessing method. The comparison results are shown in Table 1.

[0066] Table 1

[0067] The formula for calculating the anomaly detection accuracy (Acc) is as follows:

[0068] TP represents the amount of real abnormal data that was correctly identified as abnormal, TN represents the amount of real normal data that was correctly identified as normal, FP represents the amount of real normal data that was incorrectly identified as abnormal, and FN represents the amount of real abnormal data that was incorrectly identified as normal.

[0069] The formula for calculating the interpolation error RMSE is as follows: ; in, n The number of data samples. These are measured values ​​of meteorological variables such as wind speed and power. These are the interpolated values ​​for the corresponding data.

[0070] Data availability after data cleaning for:

[0071] in, This represents the effective data volume after cleaning. This represents the total amount of the original data.

[0072] As shown in Table 1, the data preprocessing scheme of the present invention has significantly higher anomaly identification accuracy and data availability than traditional methods, and significantly lower interpolation error RMSE than traditional methods, verifying the superiority of the present invention in the data preprocessing stage.

[0073] Step S2: Using historical measured wind speeds as supervised values, the wind shear power law formula is incorporated into the GRU neural network model as a physical constraint to perform localized correction on the reanalysis wind speed.

[0074] By introducing the wind shear power law formula as a physical prior, a GRU neural network model integrating physical constraints is constructed. The core formula of GRU and the total loss function are substituted into the model, and the measured wind speed is used as the supervision value to perform local correction on the reanalysis wind speed. The corrected wind speed RMSE=0.68m / s. Substituting the accuracy improvement rate formula, the accuracy of the reanalysis wind speed is calculated to be 41.5% higher than that of the uncorrected reanalysis wind speed.

[0075] The GRU correction based on the physical constraint of wind shear power law introduced in this invention is compared with ordinary GRU correction, uncorrected wind speed, and other correction methods. The comparison results are shown in Table 2.

[0076] Table 2

[0077] The formula for calculating the improvement rate of correction accuracy is as follows: ; in, For the RMSE of uncorrected reanalysis wind speed, The RMSE of the corrected wind speed.

[0078] The formula for calculating MAE is: ; in, n The number of data samples. These are measured values ​​of meteorological variables such as wind speed and power. These are the interpolated values ​​for the corresponding data.

[0079] As shown in Table 2, the corrected wind speed RMSE is ≤0.75m / s, which is ≥40% higher than the accuracy of the uncorrected reanalysis wind speed, and the correction time per station is ≤85ms. The local correction method of this invention has the lowest RMSE, the highest accuracy improvement rate, and the best physical consistency after correction, verifying the superiority of the fusion wind shear power-law physical prior GRU correction method.

[0080] Step S3: Construct a time-varying t-Copula model differentiated by month, independently fit the tail correlation parameters month by month, combine the rank correlation coefficient to explore the correlation between variables, and screen key meteorological variables.

[0081] A time-varying t-Copula model was constructed for 12 months. The time-varying t-Copula density formula and the Kendall and Spearman rank correlation coefficient formula were substituted into the model. The correlation matrix and degree of freedom parameters were fitted month by month. The key meteorological variables were calculated as wind speed (correlation coefficient 0.89), temperature (correlation coefficient -0.37), and air pressure (correlation coefficient -0.41). Redundant features were removed, and the feature dimension was reduced by 45%.

[0082] The feature selection of the time-varying t-Copula model of the present invention is compared with other feature selection methods, and the comparison results are shown in Table 3.

[0083] Table 3

[0084] As shown in Table 3, the feature selection method of this scheme has the lowest feature dimension, the highest prediction accuracy, and the best seasonal adaptability, which verifies the superiority of the time-varying t-Copula feature selection method in this embodiment of the invention.

[0085] Step S4: Fit the corrected wind speed using a Weibull distribution, integrate the wind speed interval, input the result into the improved LSTM power curve to convert it into hourly theoretical power, and sum them up to obtain the monthly power generation.

[0086] The corrected wind speed sequence was fitted with a Weibull distribution. Substituting the Weibull distribution formula and the R² calculation formula, the goodness of fit R² = 0.97 was obtained. The wind speed interval was converted into hourly power by improving the LSTM power curve formula. Substituting the power into the monthly power generation accumulation formula, the monthly power generation was obtained by accumulation. The sample size was expanded from the original 24 (2 years of measured monthly data) to 960.

[0087] Step S5: Construct a GBDT model using historical power generation data for the same month and selected key meteorological variables as input features, and introduce dynamic elastic weight regularization.

[0088] Using historical monthly power generation and key meteorological variables as inputs, a time-series driven GBDT model for the same month is constructed by substituting the GBDT prediction formula and the dynamic EWC regularized loss function. Dynamic EWC elastic weight regularization is introduced, and the prediction error change rate is calculated in real time and the regularization coefficient is adjusted. The model converges at 132 steps. Substituting these into the stability improvement rate formula, a stability improvement of 33.1% is calculated.

[0089] The present invention compares the EWC dynamic elastic weight model with the GBDT model with fixed weight and no anti-drift measures. The EWC dynamic elastic weight anti-drift method is compared with the anti-drift method. The comparison results are shown in Table 4.

[0090] Table 4

[0091] The model stability improvement rate is calculated as follows: ; in, The standard deviation of the prediction accuracy of the GBDT model without EWC regularization. The standard deviation of the prediction accuracy of the GBDT model with EWC regularization is given.

[0092] As shown in Table 4, the GBDT model based on dynamic elastic weights in this invention improves stability by more than 32% compared to the model without EWC regularization. This invention exhibits the best anti-drift performance, verifying the superiority of the dynamic EWC improvement.

[0093] Step S6: The GBDT model outputs the monthly power generation, annual total power generation, and predicted trend curves for the future preset time period.

[0094] Substituting the predicted monthly power generation into the formula for total annual power generation and the prediction accuracy formula, the system outputs the monthly power generation, total annual power generation, and prediction trend curve for the wind farm over the next 12 months, with an average prediction accuracy of 89.7%.

[0095] The prediction performance of the wind power long-term generation prediction method based on time-varying Copula-GBDT of this invention is compared with the prediction results of traditional long-term generation prediction methods. The trend curves are compared, for example... Figure 2 As shown in Table 5, the comparison results of each evaluation indicator are shown in Table 5.

[0096] Table 5

[0097] Based on Table 5, the indicators of the wind power long-term generation prediction method based on time-varying Copula-GBDT of the present invention are all superior to those of traditional methods, which fully demonstrates the technical advantages of the present invention.

[0098] This invention forms a complete long-term power generation prediction closed-loop system through intelligent cleaning and restoration of multi-source data, reanalysis of wind speed correction using GRU neural networks with fused wind shear power-law physical priors, screening of nonlinear tail correlation features of monthly time-varying t-Copula, joint power conversion using Weibull distribution and improved power curves, same-month time-series driven GBDT modeling, and dynamic elastic weight anti-concept drift calibration. This invention can stably output monthly power generation prediction results for the next 12 months, with a monthly power generation prediction accuracy of 88.55%–99.97%, an average prediction accuracy of no less than 88.5%, and a prediction accuracy of ≥85% even under extreme weather conditions. This represents an improvement of 0.94%–30.54% compared to traditional continuous methods, time-series models, and conventional machine learning methods, and can be widely used in annual wind farm power generation planning, grid renewable energy planning, power trading, and meteorological support decision-making.

[0099] The prediction time for a single station is ≤300ms, and the prediction time for the entire station in batches is ≤5000ms, making it suitable for batch applications by provincial meteorological service centers. Moreover, no new hardware equipment is required, and it can be deployed based on existing meteorological data and wind farm measured data, resulting in low engineering implementation costs and high cost-effectiveness.

[0100] Finally, it should be noted that in this document, relational terms such as "first" and "second" are used only to distinguish one entity or operation from another, and do not necessarily require or imply any such actual relationship or order between these entities or operations. Furthermore, the terms "comprising," "including," or any other variations thereof are intended to cover non-exclusive inclusion, such that a process, method, article, or apparatus that comprises a list of elements includes not only those elements but also other elements not expressly listed, or elements inherent to such a process, method, article, or apparatus. Without further limitations, an element defined by the phrase "comprising one..." does not exclude the presence of other identical elements in the process, method, article, or apparatus that includes said element.

[0101] The various embodiments in this specification are described in a progressive manner, with each embodiment focusing on the differences from other embodiments. The same or similar parts between the various embodiments can be referred to each other.

[0102] The above description of the disclosed embodiments enables those skilled in the art to make or use the invention. Various modifications to these embodiments will be readily apparent to those skilled in the art, and the general principles defined herein may be implemented in other embodiments without departing from the spirit or scope of the invention. Therefore, the invention is not to be limited to the embodiments shown herein, but is to be accorded the widest scope consistent with the principles and novel features disclosed herein.

Claims

1. A method for predicting long-term wind power generation based on time-varying Copula-GBDT, characterized in that, include: Collect measured data from wind farms and historical reanalysis meteorological data, and perform anomaly identification, missing data imputation and restoration to obtain multi-source datasets; Using historical measured wind speeds as supervised values, the wind shear power law formula is incorporated into the GRU neural network model as a physical constraint to perform localized correction on the reanalysis wind speed. We constructed a time-varying t-Copula model that distinguishes by month, independently fitted the tail correlation parameters month by month, and combined the rank correlation coefficient to explore the correlation between variables and screen key meteorological variables. The wind speed is fitted with a Weibull distribution and corrected. The wind speed interval is integrated and then input into the improved LSTM power curve to convert it into hourly theoretical power. The monthly power generation is then accumulated. A GBDT model was constructed using the power generation of the same month in previous years and the selected key meteorological variables as input features, and dynamic elastic weight regularization was introduced. The GBDT model outputs monthly power generation, annual total power generation, and predicted trend curves for a future preset time period.

2. The wind power long-term generation prediction method based on time-varying Copula-GBDT according to claim 1, characterized in that, include: The three-sigma principle, bidirectional quartile method and DBSCAN clustering are used to jointly identify abnormal data; By combining KNN with LSTM for temporal extrapolation, missing data can be imputed and restored to obtain a multi-source dataset.

3. The wind power long-term generation prediction method based on time-varying Copula-GBDT according to claim 1, characterized in that, The total loss function of the GRU neural network model is: ； in, This represents the mean square error of wind speed prediction. For physical prior weights, Output wind speed for the GRU neural network model. This is for calculating wind speed based on the power-law formula for wind shear.

4. The wind power long-term generation prediction method based on time-varying Copula-GBDT according to claim 1, characterized in that, The time-varying t-Copula model, which independently fits tail correlation parameters month by month and combines them with rank correlation coefficients to mine variable associations, includes: Based on multi-meteorological variable data from a multi-source dataset, a time-varying t-Copula model is constructed by month, and the correlation matrix and degree of freedom parameters are fitted month by month. Traverse all meteorological variables and extract the binary correlation coefficients between each set of meteorological variables and power generation from the correlation matrix; The Kendall rank correlation coefficient and Spearman rank correlation coefficient between each meteorological variable and power generation were calculated using the t-Copula function based on each binary correlation coefficient. Determine the association between variables based on the correlation coefficient threshold; Key meteorological variables were selected based on correlations.

5. The wind power long-term generation prediction method based on time-varying Copula-GBDT according to claim 4, characterized in that, The process of screening key meteorological variables includes: If the absolute value of the Kendall rank correlation coefficient of the current meteorological variable is greater than the Kendall coefficient threshold, and the absolute value of the Spearman rank correlation coefficient is greater than the Spearman coefficient threshold, then the current meteorological variable is a key meteorological variable. If the absolute value of the rank correlation coefficient of the current meteorological variable only meets the Kendall coefficient threshold or the Spearman coefficient threshold, then the meteorological variable with direct physical effect is retained as the key meteorological variable based on the physical mechanism. If the absolute value of the rank correlation coefficient of the current meteorological variable does not meet the Kendall coefficient threshold and the Spearman coefficient threshold, then the current meteorological variable will be removed.

6. The wind power long-term generation prediction method based on time-varying Copula-GBDT according to claim 4, characterized in that, The probability density formula for the t-Copula function is: ； in, As a meteorological variable dimension, For the first m The correlation matrix obtained by fitting over a month For the first m The degrees of freedom obtained by fitting over a month For the first i The empirical distribution function values ​​of the meteorological variables It follows a univariate t-distribution. Single variable t The inverse function of the distribution; The formula for Kendall's rank correlation coefficient is: ； in, This represents the binary correlation coefficient between meteorological variables and power generation within the monthly correlation matrix. The Kendall rank correlation coefficient between the two. The formula for the Spearman rank correlation coefficient is: 。 7. The wind power long-term generation prediction method based on time-varying Copula-GBDT according to claim 1, characterized in that, Wind speed is converted into hourly theoretical power based on an improved LSTM power curve: ； in, For LSTM model parameters, This refers to the rated power of the fan. This is the corrected wind speed; when the wind speed is lower than the cut-in wind speed or higher than the cut-out wind speed. ; The monthly power generation is obtained by summing them up: ； Where N is the number of hours in the current month. for t Wind speed after constant correction For time intervals.

8. The wind power long-term generation prediction method based on time-varying Copula-GBDT according to claim 1, characterized in that, The prediction formula for the GBDT model is: ； in, For the first m The predicted values ​​of each decision tree. For learning rate, For the first m The output of each decision tree; The calculation process for the dynamic elastic weight regularization coefficient is as follows: ； in, As the benchmark coefficient, For adaptive coefficients, For the rate of change of prediction error; The dynamic elastic weight regularization loss function is: ； in, For cross-entropy loss, For dynamic regularization coefficients, For the Fisher information matrix i One element, These are the current model parameters. These are the historically optimal model parameters.

Images